A useful mental model
When you give an LLM a sentence, it predicts what token or sequence is likely to come next based on patterns learned during training and the context in the request. It can produce remarkably useful writing and code, but a plausible continuation is not the same thing as a checked database lookup.
What an LLM can do well
- Draft and revise text.
- Summarize and transform documents.
- Extract fields into a requested format.
- Explain code or concepts.
- Generate options for a human to review.
Where an LLM needs help
Current facts, private company knowledge, exact calculations, permissions, and external actions usually need retrieval, tools, code, or a human check. The model can write the answer while another system supplies the evidence and enforces the boundary.
The practical question
Do not ask whether an LLM is intelligent enough in the abstract. Ask whether it can produce an approved result for this task, with this input, at this cost, and with an acceptable failure mode.